Local-first context for Cursor and Claude with Context Harness

If you use Cursor or Claude with MCP (Model Context Protocol), you can give the AI access to your docs, Git repos, and runbooks. The catch: many setups send that context through third-party APIs or only work with a single source. I wanted one place to index everything and query it locally—no cloud dependency for the data itself.

Context Harness is a small Rust CLI and HTTP server that does exactly that. You point it at folders, Git repos, or S3 prefixes; it ingests into a local SQLite database, optionally embeds with a local model (or Ollama/OpenAI), and exposes search and retrieval over MCP. Cursor and Claude connect to it as an MCP server and get search and get tools over your indexed context.

Why local-first?

• Privacy: Docs and code never leave your machine unless you choose an external embedding API.
• Offline: With local embeddings (fastembed or tract), you can index and search without any network after the first model download.
• One config: Same TOML and connectors whether you use local, Ollama, or OpenAI for embeddings.

What you get

• Connectors: Filesystem, Git, S3, and Lua scripts for custom sources (e.g. Jira, GitHub Issues).
• Storage: SQLite with FTS5 for keyword search; optional vector embeddings for semantic (or hybrid) search.
• MCP server: One HTTP endpoint; add it to Cursor’s MCP config and the AI can search and retrieve from your knowledge base.
• CLI: ctx init, ctx sync all, ctx search "query", ctx serve mcp.

Quick start

1. Download the latest release for your platform (macOS, Linux, Windows).
2. Copy the example config and add a connector (e.g. a Git repo or a folder of markdown).
3. Run ctx init, ctx sync all, then ctx serve mcp.
4. In Cursor, add the server URL to .cursor/mcp.json and restart.

The docs have a full quick start, configuration reference, and guides for Cursor, RAG, and multi-repo setups.

Local embeddings on every platform

Recent releases ship local embeddings on all six targets (Linux glibc/musl, macOS Intel/Apple Silicon, Windows). No ONNX Runtime install—primary platforms use fastembed with bundled ORT; musl and Intel Mac use a pure-Rust tract backend. Same provider = "local" in config; models download on first use.

If you’re already using MCP with Cursor or Claude and have a pile of internal docs or code, Context Harness might be worth a try. It’s AGPL-3.0 licensed and the repo has a demo you can click through without installing anything.

Repo: github.com/parallax-labs/context-harness
Docs: parallax-labs.github.io/context-harness